AI detectors flag human writing from before ChatGPT as mostly machine

In spring 2026, a writer and AI startup founder decided to test whether AI-detection tools could distinguish her own writing from AI-generated text. She selected eight compositions spanning two decades: four pre-ChatGPT (a 2019 personal setup guide, a 2020 M.S. thesis on digital privacy, a pre-2019 essay on social media disinformation, and a 2020 grad school application essay) and four from the modern era (a science novella about mass in the Big Bang, multiple chapters of vampire fiction, an essay about her AI startup, and spiritual writing). She then ran each through eleven detectors: Originality, Ace, Humalingo, ZeroGPT, Grammarly, GPTZero, Getsolved, Quillbot, Pangram, CopyLeaks, and WinstonAI.

The results were catastrophic for the detectors' credibility. Her 2019 setup guide—provably human, written before the current AI discourse—scored 5% human according to Getsolved, 16% according to Humalingo, and 16% according to Ace. Yet six other tools rated it 100% human. This is not disagreement at the margins; it is a 95-point spread on identical words. Her 2020 grad school essay scored 15% human; her pre-2019 social media essay scored 12%. Three detectors (Ace, Humalingo, and Getsolved) failed the control group entirely, rating all eight pieces as majority or near-entirely AI, including everything predating ChatGPT.

The irony deepened when she tested her most AI-assisted work: a serialized fiction project about physics and vampires, where Claude serves as a continuity department, maintaining character sheets, chronology, and research. This piece—the one closest to machine assistance—scored 99–100% human on seven of eleven detectors, 91% on an eighth, and achieved the highest average humanity score of all eight texts tested. It outscored even her pre-LLM archive. Getsolved rated it 60–73% human, while scoring her pre-LLM writings 5–15%. One detector had inverted the truth entirely.

The writer argues the detectors are not measuring authorship but predictability. Prose with high lexical density and strong syntactic consistency mirrors the output of low-temperature language models—a feature of science and technical writing, not a sign of AI origin. Her vampire fiction succeeds because it contains unpredictable narrative choices (gravity styled as a mob boss taking stage direction from a character named Silvio Dante; the strong nuclear force personified as a seductress) that fall outside the detectors' training distributions. The tools flag her particular voice, cultivated since her 2004 Blogger days, as suspicious regardless of era.

Meanwhile, major science fiction publications have adopted AI-detection policies. Clarkesworld (July 2026) announced it will not consider submissions written, developed, or assisted by AI tools. Asimov's Science Fiction (July 2026) issued a similar policy with higher stakes: attempted submission of AI-assisted works may result in permanent bans. The writer notes the irony—Asimov's Foundation novels, set across millennia, depict AI and advanced technology everywhere. She argues that institutions using these tools to accuse students, reject submissions, or damage reputations should "take several seats," and pledges to change her mind only if the eleven detectors agreed with one another and passed her control group. They have not.

The core problem is that AI-detection tools are not measuring authorship—they are measuring predictability and stylistic consistency. The writer's controlled experiment exposes a fundamental flaw: three detectors (Ace, Humalingo, and Getsolved) flagged all eight of her human-written pieces, including texts from 2004–2020 that predate ChatGPT by years, as majority or near-entirely AI-generated. This is not a marginal disagreement; it is a categorical failure. Two thermometers reading the same winter day as 5 and 100 degrees do not both measure temperature—one is broken and useless.

The detectors appear to flag her writing because of its structural features: high lexical density, strong syntactic consistency, and information precision. These are hallmarks of her deliberate, systems-thinking voice—and they happen to mirror the output of language models running at low temperature, where less probable tokens are suppressed. But low-temperature sampling is not authorship; it is one technical mechanism among many. Her most AI-assisted piece, where Claude handled filing, continuity, and research, scored as more human than her pre-LLM archive on most detectors, because it contained unpredictable narrative choices (gravity as a mob boss, the strong nuclear force as a seductress) that fall outside the detectors' training distributions.

The stakes are concrete and immediate: Clarkesworld and Asimov's Science Fiction both reject AI-assisted work, with permanent bans for attempted submissions. Teachers are using these tools to accuse students of fraud. Editors and publishers are rejecting submissions. The writer's evidence suggests these institutions are deploying instruments that cannot agree with each other and that systematically misclassify human writing created before the technology existed.

No comments yet. Be the first to share your thoughts!